However, they hardly compare against the conventional next-token-prediction objective, which suggests they don't beat it, except for highly non-linear tasks like sudoku. Unfortunately, machine learning papers often tend to focus exclusively on the positives without mentioning bad aspects.

This makes me wonder what these people think an "LLM" even is. The name itself implies that it is large and about language. It does not imply that an "LLM" is inherently autoregressive or that it has any particular architecture even though most LLMs are transformers. Anything can be a large language model as long as it is large and models language.

NCM is basically a hierarchical diffusion transformer where the further layers take one big token and then generate two smaller tokens. So it is more like an evaluation strategy. You could apply this to any model architecture.

I personally think the name NCM itself is terrible because it contains the word "next" which inherently implies an ordering and somehow we are supposed to think that "LLM" which contains no mention of ordering, is constrained to ordered token production.